Superpowers: Forcing AI Coding Agents to Write Testable Code via TDD & Sub-Agents

The Problem: AI Agents Cut Corners

The author asked Claude Code to add a CSV export feature. The agent delivered working code and passing tests in under a minute, but two days later the feature failed in production because the tests only covered ideal inputs — no escaping for commas inside fields. This mirrors a recurring pattern: AI agents write tests that merely rubber-stamp their own implementation, skipping edge cases and engineering rigor.

What Superpowers Does

Superpowers (GitHub: https://github.com/obra/superpowers) is not a model upgrade or a prompt pack. It is a collection of 15+ independent skill modules — each with its own trigger logic — that are injected into the agent's context at every new session via a SessionStart hook. The skills cover brainstorming, writing implementation plans, test-driven development, systematic debugging, sub-agent delegation, and two-phase code review. Because the skills are presented as mandatory context, the agent cannot opt out; the workflow forces compliance.

Core Enforcement: Test-Driven Development Skill

The test-driven-development skill mandates the full red–green–refactor cycle:

Write a failing test first.

Watch it run red.

Write the minimal code to make it green.

Only then refactor.

If the agent writes implementation code before a test, the skill requires it to delete that code and start over. This reverses the usual AI habit of writing tests that mirror the (possibly buggy) implementation, turning tests into a specification that the code must satisfy.

End-to-End Workflow

When a user proposes an idea, the flow proceeds as follows:

Brainstorming skill asks clarifying questions and produces a design doc for step-by-step confirmation.

Writing-plans breaks the design into small tasks (file paths, verification steps, each doable in minutes).

A new git worktree is created on a fresh branch, isolating the work.

Subagent-driven-development dispatches each task to a brand-new child agent that receives only the task spec and relevant code — no historical context baggage.

Two-phase review : first, checklist against the plan for completeness; second, code-quality gate. Severe issues are rejected immediately.

Only after all tasks pass does the agent propose a PR, direct merge, or further iteration.

The author notes that with this pipeline, agents can work for 1–2 hours without drifting, whereas long single-session contexts previously led to repeated mistakes.

Installation Across Tools

Superpowers supports multiple AI coding environments:

/plugin install superpowers@claude-plugins-official

For the author's own marketplace (faster updates):

/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace

Cursor: /add-plugin superpowers Gemini CLI:

gemini extensions install https://github.com/obra/superpowers

Codex, OpenCode, Kimi Code are also supported. No extra configuration is needed; skills auto-trigger.

Chinese Localization

A community fork superpowers-zh (https://github.com/jnMetaCode/superpowers-zh) provides full Chinese translation of all skills plus a few localized additions, installable via a single npx command into 20+ tools (MIT licensed).

When to Use (and Not Use) Superpowers

The author explicitly states Superpowers is not for throwaway scripts or quick prototypes — the overhead of brainstorming, planning, and review takes 10+ minutes for maybe 10 lines of code. It shines for code that will be maintained long-term, handed off to others, or must be auditable. Trade-offs include higher token consumption (each sub-agent rebuilds context) and a known issue where session compression on some platforms (e.g., Hermes Agent) can drop the skill injection; the workaround is to start a new session.

Author's Take

The author, historically skeptical of "AI gets smarter" tools, values Superpowers precisely because it does not claim to boost intelligence. Instead, it enforces discipline through process — something human teams struggle with, let alone agents with no KPI or shame. The verdict: "It doesn't make the agent smarter; it just stops it from being lazy. That's worth the stars."

Additional reference: https://github.com/superpowers-ai/superpower